Anveld Docs
Platform

Training & Simulation

The local-first workflow for simulator-driven iteration, distributed training, and experiment tracking.

Simulation is not treated as a sidecar in Anveld. It is the primary development environment.

Local-First by Default

The intended workflow is:

  1. Implement your robot or agent locally.
  2. Run in simulation with anveld dev.
  3. Train with anveld train.
  4. Move the same project toward remote compute or deployment targets when needed.

Infrastructure in Scope

Anveld includes platform support for:

  • Ray-backed local and distributed execution
  • Experiment tracking and comparison
  • Simulator backends through plugins
  • Environment wrappers and parallel simulation utilities
  • Domain randomization helpers

Simulator Integrations

The repository already includes official simulator plugins such as:

  • MuJoCo
  • Isaac Sim

Because they are plugins, teams can swap environment implementations without forking the SDK core.

Why It Is Structured This Way

Robotics teams usually pay a tax for stitching together simulation, training, logging, and deployment as separate systems. Anveld's platform layer exists to reduce that glue code while keeping the plugin boundary clear.

On this page